mturac/simulacra

Simulacra is a skill for AI coding assistants that reviews code by pretending to be several different fictional users trying to break it. Instead of returning a plain list of lint style warnings, it generates a small cast of ghost personas, each one hunting for a different kind of problem, and has them narrate a short story of exactly how your code would fail in their hands. Three ghosts run by default, with up to seven available on request. The Adversary looks for things like race conditions and back button abuse. Enterprise Edge looks for business logic gaps such as multi currency handling or tax rules. Chaos Monkey simulates infrastructure problems like timeouts and partial failures. Other optional ghosts focus on readability for future maintainers, performance at much larger scale, security and compliance issues, and long term maintenance risk. Each ghost produces a timeline of what it did, a description of what broke, and a bug report with a severity and suggested fix, plus a machine readable version for feeding into other tools. A built in complexity check skips this whole process for trivial changes, so simple code is not forced through unnecessary theater. The project also includes a second, separate skill called Skill Scorer, which rates any skill instruction file on a 0 to 100 scale across ten dimensions and gives a clear verdict of ship it, rework, or scrap. Simulacra installs as a skill or plugin into several AI coding tools, including Claude Code, Codex CLI, Cursor, Windsurf, and others, and is built entirely from plain markdown files with no external dependencies or build step. According to the project's own benchmark against 25 code snippets with known bugs, it correctly found about 83 percent of the seeded bugs, with a false positive rate under 4 percent. It is released under the MIT license.